lecrapaud 0.21.2

Framework for machine and deep learning, with regression, classification and time series analysis

Welcome to LeCrapaud

An all-in-one machine learning framework

🚀 Introduction

LeCrapaud is a high-level Python library for end-to-end machine learning workflows on tabular data, with a focus on financial and stock datasets. It provides a simple API to handle feature engineering, model selection, training, and prediction, all in a reproducible and modular way.

✨ Key Features

• 🧩 Modular pipeline: Feature engineering, preprocessing, selection, and modeling as independent steps
• 🤖 Automated model selection and hyperparameter optimization
• 📊 Easy integration with pandas DataFrames
• 🔬 Supports both regression and classification tasks
• 🛠️ Simple API for both full pipeline and step-by-step usage
• 📦 Ready for production and research workflows

⚡ Quick Start

Install the package

pip install lecrapaud

How it works

This package provides a high-level API to manage experiments for feature engineering, model selection, and prediction on tabular data (e.g. stock data).

Typical workflow

from lecrapaud import LeCrapaud

# 1. Create the main app
app = LeCrapaud(uri=uri)

# 2. Define your experiment context (see your notebook or api.py for all options)
context = {
    "data": your_dataframe,
    "columns_drop": [...],
    "columns_date": [...],
    # ... other config options
}

# 3. Create an experiment
experiment = app.create_experiment(**context)

# 4. Run the full training pipeline
experiment.train(your_dataframe)

# 5. Make predictions on new data
predictions = experiment.predict(new_data)

Database Configuration (Required)

LeCrapaud requires access to a MySQL database to store experiments and results. You must either:

• Pass a valid MySQL URI to the LeCrapaud constructor:

  app = LeCrapaud(uri="mysql+pymysql://user:password@host:port/dbname")
  

• OR set the following environment variables before using the package:
  • DB_USER, DB_PASSWORD, DB_HOST, DB_PORT, DB_NAME
  • Or set DB_URI directly with your full connection string.

If neither is provided, database operations will not work.

Using OpenAI Embeddings (Optional)

If you want to use the columns_pca embedding feature (for advanced feature engineering), you must set the OPENAI_API_KEY environment variable with your OpenAI API key:

export OPENAI_API_KEY=sk-...

If this variable is not set, features relying on OpenAI embeddings will not be available.

Experiment Context Arguments

The experiment context is a dictionary containing all configuration parameters for your ML pipeline. Parameters are stored in the experiment's database record and automatically retrieved when loading an existing experiment.

Required Parameters

Parameter	Type	Description	Example
data	DataFrame	Input dataset (required for new experiments only)	pd.DataFrame(...)
experiment_name	str	Unique name for the experiment	'stock_prediction'
date_column	str	Name of the date column (required for time series)	'DATE'
group_column	str	Name of the group column (required for panel data)	'STOCK'

Feature Engineering Parameters

Parameter	Type	Default	Description
columns_drop	list	[]	Columns to drop during feature engineering
columns_boolean	list	[]	Columns to convert to boolean features
columns_date	list	[]	Date columns for cyclic encoding
columns_te_groupby	list	[]	Groupby columns for target encoding
columns_te_target	list	[]	Target columns for target encoding

Preprocessing Parameters

Parameter	Type	Default	Description
time_series	bool	False	Whether data is time series
val_size	float	0.2	Validation set size (fraction)
test_size	float	0.2	Test set size (fraction)
columns_pca	list	[]	Columns for PCA transformation
pca_temporal	list	[]	Temporal PCA config (e.g., lag features)
pca_cross_sectional	list	[]	Cross-sectional PCA config (e.g., market regime)
columns_onehot	list	[]	Columns for one-hot encoding
columns_binary	list	[]	Columns for binary encoding
columns_ordinal	list	[]	Columns for ordinal encoding
columns_frequency	list	[]	Columns for frequency encoding

Feature Selection Parameters

Parameter	Type	Default	Description
percentile	float	20	Percentage of features to keep per selection method
corr_threshold	float	80	Maximum correlation threshold (%) between features
max_features	int	50	Maximum number of final features
max_p_value_categorical	float	0.05	Maximum p-value for categorical feature selection (Chi2)

Model Selection Parameters

Parameter	Type	Default	Description
target_numbers	list	[]	List of target indices to predict
target_clf	list	[]	Classification target indices
models_idx	list	[]	Model indices or names to use (e.g., [1, 'xgb', 'lgb'])
max_timesteps	int	120	Maximum timesteps for recurrent models
perform_hyperopt	bool	True	Whether to perform hyperparameter optimization
number_of_trials	int	20	Number of hyperopt trials
perform_crossval	bool	False	Whether to use cross-validation during hyperopt
plot	bool	True	Whether to generate plots
preserve_model	bool	True	Whether to save the best model
target_clf_thresholds	dict	{}	Classification thresholds per target

Example Context Configuration

context = {
    # Required parameters
    "experiment_name": f"stock_{datetime.now().strftime('%Y%m%d_%H%M%S')}",
    "date_column": "DATE",
    "group_column": "STOCK",
    
    # Feature selection
    "corr_threshold": 80,
    "max_features": 20,
    "percentile": 20,
    "max_p_value_categorical": 0.05,
    
    # Feature engineering
    "columns_drop": ["SECURITY", "ISIN", "ID"],
    "columns_boolean": [],
    "columns_date": ["DATE"],
    "columns_te_groupby": [["SECTOR", "DATE"]],
    "columns_te_target": ["RET", "VOLUME"],
    
    # Preprocessing
    "time_series": True,
    "val_size": 0.2,
    "test_size": 0.2,
    "pca_temporal": [
        # Old format (still supported)
        # {"name": "LAST_20_RET", "columns": [f"RET_-{i}" for i in range(1, 21)]},
        # New simplified format - automatically creates lag columns
        {"name": "LAST_20_RET", "column": "RET", "lags": 20},
        {"name": "LAST_10_VOL", "column": "VOLUME", "lags": 10},
    ],
    "pca_cross_sectional": [
        {
            "name": "MARKET_REGIME",
            "index": "DATE",
            "columns": "STOCK",
            "value": "RET",
        }
    ],
    "columns_onehot": ["BUY_SIGNAL"],
    "columns_binary": ["SECTOR", "LOCATION"],
    "columns_ordinal": ["STOCK"],
    
    # Model selection
    "target_numbers": [1, 2, 3],
    "target_clf": [1],
    "models_idx": ["xgb", "lgb", "catboost"],
    "max_timesteps": 120,
    "perform_hyperopt": True,
    "number_of_trials": 50,
    "perform_crossval": True,
    "plot": True,
    "preserve_model": True,
    "target_clf_thresholds": {1: {"precision": 0.80}},
}

# Create experiment
experiment = app.create_experiment(data=your_dataframe, **context)

Important Notes

1. Context Persistence: All context parameters are saved in the database when creating an experiment and automatically restored when loading it.

2. Parameter Precedence: When loading an existing experiment, the stored context takes precedence over any parameters passed to the constructor.

3. PCA Time Series:

   • For time series data, both pca_cross_sectional and pca_temporal automatically use an expanding window approach with periodic refresh (default: every 90 days) to prevent data leakage.
   • The system fits PCA only on historical data (lookback window of 365 days by default) and avoids look-ahead bias.
   • For panel data (e.g., multiple stocks), lag features are created per group when using the simplified pca_temporal format.
   • Missing PCA values are handled with forward-fill followed by zero-fill to ensure compatibility with downstream models.

4. PCA Temporal Simplified Format:

   • Instead of manually listing lag columns: {"name": "LAST_20_RET", "columns": ["RET_-1", "RET_-2", ..., "RET_-20"]}
   • Use the simplified format: {"name": "LAST_20_RET", "column": "RET", "lags": 20}
   • The system automatically creates the lag columns, handling panel data correctly with group_column.

5. OpenAI Embeddings: If using columns_pca with text columns, ensure OPENAI_API_KEY is set as an environment variable.

6. Model Indices: The models_idx parameter accepts both integer indices and string names (e.g., 'xgb', 'lgb', 'catboost').

Modular usage

You can also use each step independently:

data_eng = experiment.feature_engineering(data)
train, val, test = experiment.preprocess_feature(data_eng)
features = experiment.feature_selection(train)
std_data, reshaped_data = experiment.preprocess_model(train, val, test)
experiment.model_selection(std_data, reshaped_data)

⚠️ Using Alembic in Your Project (Important for Integrators)

If you use Alembic for migrations in your own project and you share the same database with LeCrapaud, you must ensure that Alembic does not attempt to drop or modify LeCrapaud tables (those prefixed with {LECRAPAUD_TABLE_PREFIX}_).

By default, Alembic's autogenerate feature will propose to drop any table that exists in the database but is not present in your project's models. To prevent this, add the following filter to your env.py:

def include_object(object, name, type_, reflected, compare_to):
    if type_ == "table" and name.startswith(f"{LECRAPAUD_TABLE_PREFIX}_"):
        return False  # Ignore LeCrapaud tables
    return True

context.configure(
    # ... other options ...
    include_object=include_object,
)

This will ensure that Alembic ignores all tables created by LeCrapaud when generating migrations for your own project.

---

🤝 Contributing

Reminders for Github usage

1. Creating Github repository

$ brew install gh
$ gh auth login
$ gh repo create

2. Initializing git and first commit to distant repository

$ git init
$ git add .
$ git commit -m 'first commit'
$ git remote add origin <YOUR_REPO_URL>
$ git push -u origin master

3. Use conventional commits
   https://www.conventionalcommits.org/en/v1.0.0/#summary

4. Create environment

$ pip install virtualenv
$ python -m venv .venv
$ source .venv/bin/activate

5. Install dependencies

$ make install

6. Deactivate virtualenv (if needed)

$ deactivate

---

Pierre Gallet © 2025